I think it is a difficult problem. Much of the difficulty is just framing the question in a philosophical sense.<p>The article talks about music recommendations based on a user's previous music-listening habits but what about recommendations based NOT on previous listening habits? I guess it depends on if the point of music recommendations is to "drill down" on a particular genre or related sub-genre.<p>In general music recommendations can also be made based on interests in literature or aesthetics or even historical topics. They can also be based on absence of experience-- how do you know if you've optimized globally if you haven't seen the spectrum of possibilities?<p>Finally, there's the question of what constitutes "success". Is it possible for someone to appreciate a recommendation even if they don't like it (as in "thumbs up" like it)? Aesthetic tastes are far more complex than like/dis-like.<p>I realize that these considerations aren't necessarily amenable to a pure computational approach. Perhaps the answer lies in using human decisions and computation together, like last.fm scrobbling?